Steel foundry defect analysis drops prompt rules for code gates, lifting defect-location mentions from zero to 16

In steel foundry root-cause analysis, whatever must hold every time is enforced by code rather than written into instructions to the model. Previously the rule lived only in the instructions, and the model read a bare material-standard number, took a cast steel for a ductile iron and recommended a pouring temperature below the liquidus of steel, at which the metal would freeze in the ladle. Once the material was anchored by code, the confusion did not recur in the agent runs; in the same period none of sixteen archived one-shot analyses mentioned the actual defect location.

The analysis is split into four narrowly defined expert roles plus a judge that must first try to refute their opinions before writing the report, and each role may find no defect. A single model call proved unable to reason deeply about defect physics and fill a strict output structure at the same time. The judge's first version averaged opinions and gave a secondary factor such as turbulent filling the same weight as an isolated hot spot, so it was changed to name contradictions and rejected hypotheses first and write only from what survived.

Solidification and filling fields are read only inside the intersection of the place, the moment and the threshold at which a defect forms, and velocity claims are anchored at the ingates and the cavity entrance. Previously the model searched for high velocity everywhere and found the highest values in the pouring basin, the sprue and after the filter, where the running system is closed, full and pressurized and high velocity is intended. For steel a critical velocity of the order of 0.5 m/s is reported at the ingates and the front of the stream entering the mold cavity, where oxide films can be entrained; the standard porosity export shows only regions above a threshold of a few percent, so sub-threshold microshrinkage is invisible.

Claims about equipment carry three trust tiers, and only the tier confirmed on the plant's own machine may enter an operator procedure. A pattern-change procedure cited the manual page by page, but one step was the system's own recommendation, to align an axis offset in manual mode, a task that belongs to a setter and where position protection is limited. During a trial the axis traveled beyond its working range and the procedure was withdrawn and prohibited; a comparison of the manual with years of machine records then showed several functions described in the manual that had never once been triggered on this machine.

The system was developed and tested on data from a foundry producing steel castings on an automatic green-sand molding line with a sand plant and an automatic pouring unit, and it works offline on exported historical data with no connection to the manufacturing execution system or the line's control systems, on one workstation with one user, receiving data days to weeks after events. The author limits what transfers to the ability to build such a system again, namely the gates, the methods and the measurement, while the library of a plant's documentation, the knowledge cards of its equipment, the data contracts for its exports, the dictionary of its signals and alarms and reference analyses of its castings must be rebuilt at another plant, and T1 status must be verified on each machine again. The next steps the author proposes are using the plant's existing data systematically, adding six fields per mold that the pouring unit's control system already knows but does not export, and finding out where the line data are stored and whether read access is possible.

The author states the evidence base is narrow: the depth measurement rests on one reference casting, three runs with mandates and a single stored run without them, and the reference analysis was written by an LLM under the author's guidance, not by an independent human expert, so a blind comparison with independent experts on historical cases is still missing. The mandates were derived from the same casting on which they were measured, one mandate contains a term the scorecard counts, the runs were few and their inputs varied, and the improvement was measured in the isolated agent package rather than in the production tab. The scorecard counts words and headings and cannot judge the correctness of reasoning, and the citation check verifies that a quotation is verbatim but not separately that it comes from the right page, which the author considers the largest real gap. There is no regression set run on every change of model or instructions, and a change of model at the provider can shift behavior without any change in code; cooling curves and nodal data are not delivered from the simulation, the dual geometry of casting and machined part is missing for some parts, the tapping time exists only in the melt-shop system, and scrap is not recorded per mold. No formal governance of the AI system has been set up with respect to the EU Artificial Intelligence Act or an AI management system standard, no conformity assessment has been made, and the system depends on the knowledge of one person, with a handover not yet tested.